<p>Automated assessment through the analysis of facial expressions in autism can assist in early screening, providing strong support for timely intervention and contributing to the healthy development of patients. However, current deep learning models still face several challenges in the practical application of intelligent diagnosis for Autism Spectrum Disorder (ASD). First, traditional models struggle to fully capture subtle facial expression changes and local fine-grained features. Second, existing models primarily rely on manually annotated facial expression images for supervised training, often overlooking unlabeled facial expression data. Therefore, in this study, supported by the Internet of Medical Things, we propose a Contrastive Reconstruction Vision Transformer (CREViT) model for the intelligent diagnosis of ASD. The design concept of the CREViT model integrates contrastive learning, autoencoder, and Vision Transformer (ViT). Its core advantage lies in the combination of ViT and autoencoders, enabling the model to more accurately capture subtle changes in facial expression images and learn richer and higher-quality image representations. Furthermore, by leveraging contrastive learning, the CREViT model effectively reduces its dependence on labeled data while fully utilizing unlabeled data, thereby improving the model’s generalization capability. We conducted comprehensive experiments on a real-world Autism Spectrum Disorder facial expression dataset. The results reveal that the CREViT model performs exceptionally well in intelligent ASD diagnosis, significantly enhancing prediction accuracy and model generalization ability.</p>